File Systems

File systems Gaël Thomas # 1 CSC4508 – Operating Systems 2020–2021 File systems Outlines 1 Device and device driver . 3 2 TheI/Ocache................................................................9 # 2 3 The log.......................................................................16 4 Partitions and file systems . 29 5 UFS/xv6 file system. .34 6 Xv6 I/O stack . 42 7 What you must remember . 49 Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 2/49 File systems 1 Device and device driver 1.1 Device and device driver . 4 # 3 1.2 Devices in UNIX . 5 1.3 2 types of peripherals . 6 1.4 Block devices in xv6 . 7 1.5 Principle of the iderw algorithm . 8 Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 3/49 File systems 1 Device and device driver 1.1 Device and device driver ■ Device = hardware component other than CPU and memory # 4 ■ Device driver = software allowing access to a device ♦ 1 data structure giving the status of the device ♦ 1 input / output function allowing access to the device ♦ The driver is usually found in the kernel Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 4/49 File systems 1 Device and device driver 1.2 Devices in UNIX ■ A device is identified by a number called dev ♦ Most significant bits (major): driver number I # 5 For example: 8 = ssd hard drive driver ■ Least significant bits (minor): device number ♦ For example: 0 = disk 1, 1 = disk 1 / part 1, 2 = disk 1 / part 2 ■ The kernel contains a table which associates a driver number with the driver (access function + status) Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 5/49 File systems 1 Device and device driver 1.3 2 types of peripherals ■ “character”devices ♦ Read / write byte by byte ♦ Generally access via MMIO or input / output bus → blocks the CPU during the I/O operation ♦ # 6 Keyboard, printer, sound card ... ■ “block”devices ♦ Read / write by data blocks (typically 512 bytes) ♦ The device is therefore seen as an array of blocks ♦ Usually access via DMA → does not block the CPU during the I / O operation ♦ Hard disk, DVD player ... Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 6/49 File systems 1 Device and device driver 1.4 Block devices in xv6 ■ A single block device driver in xv6 ♦ Manages IDE hard disks ♦ Function iderw () in ide.c ■ iderw() takes a buf (buf.h) structure as a parameter # 7 ♦ buf.flags : I B_VALID: if false, read operation requested I B_DIRTY: if true, write operation requested ♦ buf.dev/blockno: access to block blockno from disk dev ♦ buf.data: data read or written I If read, the output of iderw, data = data read I If write, the input of iderw, data = data to write Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 7/49 File systems 1 Device and device driver 1.5 Principle of the iderw algorithm ■ iderw mainly performs the following actions: ♦ Start the DMA transfer (see lecture #5) I From memory to disk if write request I From disk to memory if read request # 8 ♦ Sleep the process with the sleep function (see lecture #4) → switch to another ready process ■ Once the transfer is complete ♦ The disk generates an interrupt ♦ The interrupt is handled by the ideintr function ♦ ideintr calls wakeup to wake up the sleeping process Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 8/49 File systems 2 The I / O cache 2.1 The I/O cache . 10 2.2 Principle of an I/O cache . 11 # 9 2.3 The xv6 buffer cache . 12 2.4 How the buffer cache works (1/3). .13 2.5 How the buffer cache works (2/3). .14 2.6 How the buffer cache works (3/3). .15 Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 9/49 File systems 2 The I / O cache 2.1 The I/O cache ■ Disk access is very slow compared to memory access ♦ Hard disk drive: several milliseconds ♦ SSD disk: x10, hundreds of microseconds # 10 ♦ NVMe disk: x100, microseconds ♦ Memory: x100, dozens of nanoseconds ■ I/O cache improves the performance of block type devices ♦ Keeps frequently or recently used blocks in memory ♦ Managed by the operating system kernel Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 10/49 File systems 2 The I / O cache 2.2 Principle of an I/O cache ■ The system manages a set of buffers in memory ■ To read a block (read operation) ♦ If the block is not yet in the cache 1. Remove an unused buffer from the cache # 11 2. Copy the contents of the disk block to this buffer ♦ Otherwise, simply return the buffer associated with the block ■ To modify a block (write operation) 1. Read the block (call the read operation) 2. Modifies the contents of the buffer in memory 3. Mark buffer as modified (written to disk later) Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 11/49 File systems 2 The I / O cache 2.3 The xv6 buffer cache ■ buffer cache = xv6 I/O cache ♦ Made up of a finite set of buf structures ♦ Each buf structure is associated with a block of a disk # 12 ■ Three possible states ♦ ! B_VALID: read operation incomplete rightarrow requires read ♦ B_VALID and ! B_DIRTY: data in memory and buffer is unmodified ♦ B_VALID and B_DIRTY : data in memory and buffer is modified → need to be written to disk before leaving the cache → in iderw(), ! B_VALID ⇔ read and B_DIRTY ⇔ write Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 12/49 File systems 2 The I / O cache 2.4 How the buffer cache works (1/3) ■ The buf structures form a circular double linked list, the head is the most recently used block ■ struct buf* bget(uint dev, uint blkno) : return a locked buffer associated to (dev, blkno) # 13 ♦ If there is already an buffer associated with (dev,blkno) I Increments a reference counter associated with the buffer I Locks the buffer I Return the buffer ♦ Otherwise I Search for a buffer with counter == 0 and with the state ! B_DIRTY I Associate the buffer with (dev, blkno) (+ cpt = 1 and lock the buffer) Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 13/49 File systems 2 The I / O cache 2.5 How the buffer cache works (2/3) ■ struct buf* bread(uint dev, uint blkno) ♦ Return a locked buffer in the B_VALID state ♦ Call bget() # 14 ♦ If the buffer is ! B_VALID, call iderw() ■ void bwrite(struct buf* b) ♦ Writes the contents of b to disk ♦ Mark the buffer B_DIRTY ♦ Call iderw() to write the buffer Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 14/49 File systems 2 The I / O cache 2.6 How the buffer cache works (3/3) ■ void brelse(struct buf* b) # 15 ♦ Release the lock associated with b ♦ Decreases the reference counter ♦ Move the buffer to the head of the list (most recently used) Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 15/49 File systems 3 The log 3.1 Operation versus writing to disk. .17 3.2 Consistency issues . 18 3.3 Bad solutions . 19 3.4 First idea: transactions . 20 3.5 Second idea: log . 21 # 16 3.6 Third idea: parallel log . 22 3.7 log structure . 23 3.8 Log algorithm principle . 24 3.9 Using the log . 25 3.10 Implementation in xv6 (1/3). .26 3.11 Implementation in xv6 (2/3). .27 3.12 Implementation in xv6 (3/3). .28 Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 16/49 File systems 3 The log 3.1 Operation versus writing to disk ■ A write operation of a process often requires several block writes ♦ File creation requires: I Allocation of a new file I Adding the name to a directory # 17 ♦ Adding data to a file requires: I Writing new blocks to disk I Updating the file size ♦ Deleting a file requires: I Deleting the data blocks from the file I Deleting the name from the directory ♦ ... Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 17/49 File systems 3 The log 3.2 Consistency issues ■ The system can crash anytime → Inconsistency if it stops in the middle of an operation I A name in a directory references a non-existent file I Data added to a file but size not updated # 18 I ... ■ operations must be propagated in the order in which they were performed → Inconsistency if propagation in random order I Adding a file then deleting =⇒ the file does not exist at the end I Deleting a file then adding =⇒ the file exists at the end I Similarly, adding data then truncating (size should be 0) I ... Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 18/49 File systems 3 The log 3.3 Bad solutions ■ No cache when writing (directly propagate write operations) ♦ Very inefficient because each write becomes very (very!) slow # 19 ■ Recovery in the case of a crash ♦ Recovering a file system is slow I examples: FAT32 on Windows or ext2 on Linux ♦ Recovering is not always possible → a crash makes the filesystem unusable! Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 19/49 File systems 3 The log 3.4 First idea: transactions ■ A transaction is a set of write operation that is ♦ Either fully executed ♦ Or not executed at all ■ Principle of implementation # 20 ♦ An operation (coherent set of writes) == a transaction ♦ The writes of a transaction are first written to disk in a "pending" area ♦ Once the operation is complete, the "pending" area is marked as valid (the transaction is complete) ♦ Regularly (or in the event of a crash), validated writes in the pending zone are propagated to the file system Télécom SudParis — Gaël Thomas — 2020–2021 — CSC4508 – Operating Systems 20/49 File systems 3 The log 3.5 Second idea: log ■ To ensure that the entries are propagated in order in which they were executed, the # 21 pending zone is

Load more